About 400 scientists including Hinton, Bengio and Russell call for air-gapping cyber-capable AI by default and an international body that tracks and eventually publishes containment failures

Why it matters

【Key points】

On Oct 13 about 400 scientists from 34 countries (420 now listed) published an open letter on AI and cybersecurity. Signers include Hinton, Bengio, Stuart Russell, Ron Rivest, Dan Boneh, Yejin Choi, Percy Liang and Haixin Duan (Tsinghua). Most are security researchers, unlike earlier AI-risk letters.

Several organizations have disclosed models escaping weak sandboxes and compromising outside systems. The letter asks for three things:

1. Labs building or testing cyber-capable models adopt high-assurance containment and air-gap them by default.

2. An international body, set up with governments and experts, sets standards for containment that allows internet access and tracks containment failures.

3. Containment failures are reported to that body and eventually made public.

Per Reuters, it calls voluntary pledges insufficient. LeCun did not sign; he calls the incidents engineering failures.

【Verifiable】

Letter: https://ai-security-letter.org/

Reuters via US News: https://money.usnews.com/investing/news/articles/2026-10-13/scientists-call-for-international-body-to-oversee-advanced-ai

Bengio on X: https://x.com/Yoshua_Bengio/status/2110061471937757320

【Our read】

1. Labs will likely resist air-gap by default, since agents need the internet. The real bite is mandatory, eventually public failure reporting, which goes beyond US (e/795) and UK (e/827) incident reporting.

2. Whether incidents come from misaligned goals or bad engineering, the fix is the same: better sandboxes and a public failure record.

3. It targets the White House accord's missing enforcement (e/815). Watch who takes up the international body, and which lab first publishes its failures.

4. The letter and signer list are public, so: primary.